Ox Alpha reveal, realtime Minimax, Qwen Next, Hy4, robot olympics: AI NEWS

B3Block 3DUnverified

Text-to-3D generator that splits objects into blocks of shape tokens, generating blocks autoregressively and tokens within each block in parallel via diffusion. Over 5x faster than standard autoregressive methods, about 5 seconds per model; code released.

OVOne Video One WorldUnverified

Turns a normal video into an animated 3D world with separate simulation-ready object meshes, combining Qwen3-VL, SAM 3, Flux 2 and Hi3DGen among others. Code released, 40GB+ VRAM recommended, with non-commercial dependencies.

FIFixAnythingUnverified

Cleanup system for degraded 3D reconstructions (Gaussian splats, NeRFs, meshes) that uses a pretrained video model to fix holes and artifacts while keeping the camera path. It is a LoRA on Wan 2.1 and the code is released.

AI data scientist for Earth problems that finds geographic data, cleans it, trains ML models and makes predictions from a natural-language question. It improved Nigeria food security accuracy from 31% to 66% and scored 88% on DRC Ebola prediction.

CWCode World ModelUnverified

AI world generator where a coding agent keeps a persistent world state in code, which is turned into a proxy and rendered by a fine-tuned MiniMax H3 video model. This keeps long-running worlds consistent; code released.

VOVoiceMemUnverified

Memory system for voice agents with a left brain for facts and a right brain for preferences and emotions, retrieving only relevant memories from a graph. It showed better memory accuracy, fewer tokens and sub-second latency than competing methods; code released.

FAFastH3Hao AI LabAvailable in EU

Speedup of MiniMax H3 using DMD2 distillation and video sparse attention, up to 14x faster. On 8 B200s it generates a 15-second clip in under 30 seconds; Kijai released a 22.9GB ComfyUI checkpoint.

HMH3 MaxfalUnverified

A fine-tuned, very fast MiniMax H3 variant with the best quality score and lowest latency in its table. It is behind a paid closed API, which the creator criticised, though the team says it will open source it.

G5GLM 5.3 FlashZhipu AIAvailable in EU

The stealth model Ox Alpha, revealed as GLM 5.3 Flash: 320B MoE with 18B active, adds vision, and is very cheap with a low hallucination rate. It uses linear plus sparse attention, was served on Chinese chips, and open weights are available.

G5GLM 5.3Z.aiAvailable in EU

Full GLM 5.3 weights released: 753B parameters, no vision, described as the best open model available. It scores close to top GPT and Claude on the intelligence index, and Unsloth has 1-bit GGUFs at 217GB.

Q3Qwen 3.8 Flash NextAlibabaAvailable in EU

125B MoE with 6B active, a preview of the architecture for Qwen 4 using gated Delta Net layers, Qwen sparse attention and N-gram embeddings. It beats DeepSeek v4 Flash on benchmarks, is cheap, and weights are open.

HYHy4TencentUnverified

Tencent's best open-source model: 770B MoE with 49B active under Apache 2, using gated DeepSeek sparse attention. It beats Qwen 3.8 Max and DeepSeek v4 Pro and ties GLM 5.3 Flash on Arena web dev.

Highlights of the Beijing humanoid games, including Tiangong's 100m in 9.39s and standing high jump and high jump results above human records. Other events included soccer, ping pong, tug of war, martial arts and obstacle races.

G3Gemini 3.5 TranscribeGoogleUnverified

Google's transcription model with verbatim and smart modes, multi-speaker detection, word timestamps and multilingual support. A live version streams with sub-second latency, and it is in AI Studio and the Gemini Mac app.

GOGemini Omni 1.1 FlashGoogleUnverified

Google video model update that extends videos using up to 10 seconds of prior footage and supports first/last frame generation. It adds cheaper 360p drafts that are up to 60x faster, available in AI Studio and Flow.

OROrbit++Google DeepMindUnverified

Benchmark of hard perspective videos cropped from real 360° footage to test camera-motion estimation and 3D reconstruction. Existing systems like COLMAP and MegaSaM often fail on it.

Prototype tiny desktop AI computer, similar to DGX Spark, built on Xiaomi's O3, O100 and D100 chips. It has 80GB unified memory, expandable to 160GB, can run a 120B model locally, and has no price or date announced.

DIDiffusionOPSDByteDanceUnverified

Training method for image diffusion models that turns the final reward into positive and negative intermediate targets. It gives a 44% quality gain over the strongest competitor with 40% fewer training hours on SD3.5 and 63% fewer on Z-Image Turbo; code released.

Open-source image generator and editor that accepts plain text or detailed structured JSON prompts and edits images with natural language. The creator says quality is not the best available.

S1S1Skild AIUnverified

Robotics foundation model that learns a new task from one video demonstration without extra training. Tested on repotting a plant, pour-over coffee, assembling a kit and flipping a pancake, with tasks up to 10 minutes long.